Where OpenAI’s LLM is headed - The direction of several companies(1)

This article was written on October 20, 2024. You can also view the Korean version.

Why LLM?

Many people are focused on generative AI, but at the heart of it all is the Large Language Model (LLM), because it is the new intelligence that will be as good as, and perhaps better than, humans. Sam Altman, CEO of OpenAI, has dubbed this era “The Intelligence Age.”¹ It’s no exaggeration to say that LLMs are at the center of this new era. In this article, I will introduce the direction of OpenAI’s LLM development. I will post about other companies in the future. There are many companies developing LLM foundation models, but I will introduce OpenAI(GPT-4), Meta(LLaMA), and Korean companies Naver(HyperClova), SKT, and KT.

OpenAI: New Intelligence, Practical AI

OpenAI launched the ChatGPT service and started the generative AI boom. And it’s still driving the evolution of LLMs. Honestly, OpenAI’s progress since November 2022, when they launched ChatGPT, has been phenomenal.

• On November 30, 2022, launched the ChatGPT service based on GPT-3.5
• March 14, 2023 GPT-4 released.
• May 13, 2024 GPT-4o released.
• June 18, 2024 GPT-4o mini released.
• September 12, 2024 GPT-o1 mini, preview released.

Right now, it seems like OpenAI is building AI in two directions: (1) new intelligence that will go beyond humans, and (2) the practicalization of that intelligence. However, in the case of OpenAI, it feels like the direction is more focused on (3) intelligence itself.

A new intelligence beyond humans

OpenAI features AGI in its mission. OpenAI’s mission is “to ensure that artificial general intelligence benefits all of humanity.”² It’s almost as if AGI already exists, right? That’s because they’re really aiming to develop AGI. In February 2023, OpenAI published a post called Plans for AGI and Beyond.³ You can see that ChatGPT is just a waypoint on the way to that destination.

AGI is general purpose AI. Whereas traditional AIs specialize in specific areas (Go, dog/cat recognition), AGIs can make comprehensive reasoning about a wide range of things, just like humans. A few years ago, there were many observations that such an AGI was far off, but now no one doubts that we will have one in the near future.

As an intermediate point on the road to AGI, GPTs have gotten remarkably better at reasoning with each version. ⁴ Just comparing the most recent models, o1 and 4o, the difference is staggering. We used to have benchmark metrics to evaluate LLMs, but now we just measure their performance on human-solvable problems. On PhD-level science questions, we see that LLMs do better than humans.